Fix DeepSeek-V4 CUTLASS DSL LoadCacheMode path

[HanHan009527]

Summary

This PR fixes a narrow CUTLASS DSL API path mismatch in the DeepSeek-V4 dequant_gather_k_cutedsl kernel.

The kernel already imports cpasync:

from cutlass.cute.nvgpu import cpasync

but the load cache mode was referenced through a different namespace:

cpasync.CopyG2SOp(cute.nvgpu.LoadCacheMode.GLOBAL)

This updates it to use the same cpasync namespace as CopyG2SOp:

cpasync.CopyG2SOp(cpasync.LoadCacheMode.GLOBAL)

The change keeps the same load-cache behavior and only updates the API reference path.

CUTLASS DSL API source

The CUTLASS DSL cpasync module exports both LoadCacheMode and CopyG2SOp:

Source: https://github.com/NVIDIA/cutlass/blob/25e252bdce504932d83f43f07c4b8cc7f9b8e2b6/python/CuTeDSL/cutlass/cute/nvgpu/cpasync/__init__.py#L17-L23

__all__ = [
    "LoadCacheMode",
    "CopyG2SOp",
    ...
]

CopyG2SOp accepts the cpasync.LoadCacheMode enum as an input and normalizes it internally:

Source: https://github.com/NVIDIA/cutlass/blob/25e252bdce504932d83f43f07c4b8cc7f9b8e2b6/python/CuTeDSL/cutlass/cute/nvgpu/cpasync/copy.py#L43-L89

class LoadCacheMode(enum.Enum):
    GLOBAL = _cute_nvgpu_ir.LoadCacheMode.global_

class CopyG2SOp(CopyOp):
    def __init__(
        self, cache_mode: LoadCacheMode_ | LoadCacheMode = LoadCacheMode_.ALWAYS
    ):
        ...

Scope

This is intentionally a single-line compatibility fix. It does not attempt to solve other DeepSeek-V4/CUTLASS DSL startup issues, such as broader nvidia-cutlass-dsl API/package compatibility problems.

Testing

Not run in this PR branch. This change is limited to the symbol path used when constructing the cpasync.CopyG2SOp load atom.